Large Language Models (LLMs) have become powerful tools for text analysis, idea generation, and explaining complex processes. But can they solve one of the most sensitive and mission‑critical challenges in retail, manufacturing, and logistics — demand forecasting?

To answer this honestly, we ran a simple experiment: we asked Gemini, one of Google’s most advanced models, this question directly. Before that, we provided it with our article Forecasting in the Age of Market Shocks so the model could respond with context rather than in isolation.

Here’s what happened.

Gemini didn’t pretend to be omniscient:

“Black swans are the Achilles’ heel of any mathematical model, and I’m no exception. I cannot predict their date or scale. But once the swan has already landed, I react better than classical models.”

This is a crucial point: the LLM openly acknowledges that it cannot predict rare, sudden, structural events — but it can be useful in analyzing the consequences of shocks that have already occurred.

Gemini listed its own limitations — and they are fundamental for high‑precision forecasting:

• No market intuition: the model doesn’t sense the market; it simply continues statistical patterns.
• Hallucinations: under chaotic conditions, it may generate false correlations or nonexistent analogies.
• Data freshness constraints: the model doesn’t know what happened this morning unless it appears in search tools.

LLMs are analytical engines, not forecasting engines. They operate on linguistic patterns, while accurate demand prediction requires classical machine learning (ML) — models trained on your numerical history and optimized for specific accuracy metrics.

Gemini summarized it with striking clarity:

“If you have a clean sales table for 3 years and need a forecast for tomorrow with 1% accuracy — use classical ML. I will be an expensive and less accurate add‑on.”

This aligns perfectly with what we observe in production systems:

• classical ML models deliver stable, reproducible, explainable forecasts;
• LLMs serve as analytical assistants, not precise predictors.

LLMs are excellent companions — but not replacements — for ML models in operational demand forecasting.

Where LLMs add value:

• analyzing historical anomalies;
• explaining model behavior;
• generating hypotheses;
• working with textual factors (news, descriptions, reviews);
• scenario analysis after an event has occurred.

Where LLMs are useless or harmful:

• next‑day forecasts requiring percent‑level accuracy;
• automated replenishment;
• inventory management;
• high‑sensitivity KPIs (out‑of‑stock, waste, turnover).

What consistently works in real operations:

• disciplined data processes;
• correct problem formulation;
• proven ML models trained on your historical data;
• continuous validation and quality monitoring.

LLMs are tools for analysis, not for precise demand forecasting. They enhance ML workflows but do not replace them.

If your business needs forecasts reliable enough for procurement, logistics, and inventory management, classical ML models remain the right choice — while LLMs serve as an intelligent layer for explanations, diagnostics, and scenario analysis.